Fix issue in prefetching column major matrix.

[chengjunlu]

The prefetching lowering uses the incorrect shape sizes to get the tiling shape for column major matrix.

[Copilot]

Pull Request Overview

This PR fixes the tiling-shape computation for column-major matrices in the prefetch lowering by swapping the dimensions and updating the tensor type; it also adds new row-major prefetch tests.

• Swap the tensor shape dimensions for column-major support and recreate the tensor type
• Add MLIR tests covering scalar-mask and block-pointer prefetch in row-major mode

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
third_party/intel/lib/TritonIntelGPUToLLVM/LoadStoreOpToLLVM.cpp	Swap dimensions and reconstruct tensorType for column-major matrices
test/TritonIntelGPU/prefetch-to-llvm.mlir	Add new test cases for row-major ttig.prefetch scenarios

Comments suppressed due to low confidence (1)

test/TritonIntelGPU/prefetch-to-llvm.mlir:266

• The new tests cover row_major prefetch paths but lack a column_major case. Add a test with ttig.block_io = "column_major" to verify the column-major fix.

}

[etiotto]

@chengjunlu pls add an attached issue to the PR.

[etiotto]

@chengjunlu any impact on the benchmarks ?

[whitneywhtsang]

@chengjunlu any impact on the benchmarks ?

CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/16122585980

[whitneywhtsang]

Regression on GEMM (A^t@B): https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/16122585980/job/45491842774
GEMM (A^t@B) passes with second attempt, but there may be some issues of this change with that benchmark, as the same change caused failures before with the same benchmark: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/14777632147/job/41503097292.

[etiotto]

Regression on GEMM (A^t@B): https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/16122585980/job/45491842774 GEMM (A^t@B) passes with second attempt, but there may be some issues of this change with that benchmark, as the same change caused failures before with the same benchmark: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/14777632147/job/41503097292.

Thanks for the benchmark run @whitneywhtsang, looks like the code generated for the 2D block prefetch is incorrect.

Failure in unit tests

[etiotto]

Waiting for a PVC 1550 to run the benchmark tests again.

[whitneywhtsang]

Benchmark BMG CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/16184418309
It fails with GEMM (A^t@B).

[whitneywhtsang]

Fixed the error shown in #4611 (comment) in main branch and rebased this PR.
New Benchmark BMG CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/16202126345

[chengjunlu]

Rerun the failed tests which is caused by test of interpreter. The changes has no impact to the Triton Interpreter.

[chengjunlu]

Created an upstream PR triton-lang/triton#7470 for the Interpreter failure.

[etiotto]

Created an upstream PR triton-lang/triton#7470 for the Interpreter failure.

That PR is merged in and in our latest main branch. Rebasing and trying CI again.

[whitneywhtsang]

Depends on #4690

[whitneywhtsang]

Benchmark PVC 1100 CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/16228680948

[chengjunlu]

There are improvements on the flex attention kernel with the fixes.

No regression on GEMM performance.

The result is positive.